Prompt Request

Prompt Request Pipeline

Issue を書くだけで、自動実装 → PR → レビュー → マージまで全自動で回るパイプライン。

Overview

[auto] Issue 起票
  → 🚀 実装開始コメント
  → Phase A: タスク分解（Omega-bridge or Issue本文）
  → 📋 分析完了コメント
  → Phase B: 実装 + テスト
  → ✅ 実装完了コメント
  → Phase C: commit → push → PR作成
  → 🔗 PR作成コメント
  → 自動レビュー → 自動マージ → Issue close

Prerequisites

• OpenClaw Gateway running with hooks enabled
• GitHub CLI (gh) authenticated
• Git SSH access to target repository
• GitHub Webhook pointing to OpenClaw hooks endpoint

Setup

1. Register GitHub Webhook

On your GitHub repo → Settings → Webhooks → Add webhook:

• Payload URL: https://<your-openclaw-endpoint>/hooks/github
• Content type: application/json
• Secret: Your OpenClaw hooks token
• Events: Select individual events:
  • Issues
  • Pull requests
  • Pull request reviews
  • Check runs
  • Issue comments
  • Push

2. Configure OpenClaw hooks

Add this to your openclaw.json under hooks.mappings:

{
  "match": { "path": "github" },
  "action": "agent",
  "name": "GitHub",
  "sessionKey": "hook:github:{{repository.name}}:{{headers.x-github-event}}:{{issue.number}}{{pull_request.number}}{{check_run.id}}",
  "messageTemplate": "<see templates/messageTemplate.txt>",
  "deliver": true,
  "allowUnsafeExternalContent": true,
  "channel": "telegram",
  "to": "<your-chat-id>",
  "model": "anthropic/claude-opus-4-6",
  "thinking": "high",
  "timeoutSeconds": 900
}

3. Set working directory

In the messageTemplate, replace the working directory path:

• WORKDIR variable: where repositories are cloned (e.g., C:\Users\you\Dev or /home/you/dev)

4. (Optional) Omega-bridge

If you have Miyabi's omega-bridge for SWML-based task decomposition:

• Set the path to omega-bridge.ts in the messageTemplate
• If not available, the pipeline falls back to implementing directly from Issue body

Usage

Basic: Create an [auto] Issue

Title: [auto] Add utility function X

Body:
## Requirements
- Create scripts/x.sh with function do_x()
- Add tests in tests/test-x.sh

## Acceptance Criteria
- Function returns expected output
- Tests pass

Advanced: With agent personality

Title: [auto] [content] Write article about Y

Body:
Read agents/content-agent/AGENTS.md and SOUL.md first.
Follow the rules defined there.

## Topic
...

## Output
- File: articles/y.md
- Word count: 6000

Advanced: With skill reference

Title: [auto] Generate report with weather data

Body:
Read skills/weather/SKILL.md for API usage.

## Requirements
...

How It Works

Issue Events (action=opened)

1. Early exit check: If action is closed/labeled/etc → 1-line reply, stop
2. [auto] check: Title starts with [auto] or body contains <!-- auto-implement -->
3. Phase A: Task decomposition (omega-bridge or direct)
4. Phase B: Implementation (branch, code, test)
5. Phase C: Integration (commit, push, PR)
6. Progress comments posted at each phase

PR Events (action=opened/synchronize)

1. Skip bot senders (loop prevention)
2. Diff review for quality/security
3. Auto-merge if ALL conditions met:
   • PR title contains [auto] or branch starts with feature/issue-
   • Review is LGTM
   • CI checks pass (or empty = pass)
   • No merge conflicts

Safety Rules

• Never force push
• Never push directly to main
• Never run permission commands (icacls/chmod/chown)
• Max 3 CI fix retries per PR
• Bot sender events are skipped

Configuration Options

Performance (measured)

Tips

• Keep Issues small: 1 Issue = 1 clear deliverable, ≤300 lines of diff
• Be specific: The quality of the Issue body directly determines output quality
• Use templates: Create Issue templates for recurring task types
• Reference skills: Point the agent to relevant SKILL.md files for domain knowledge
• Reference agent definitions: Store AGENTS.md/SOUL.md in the repo for consistent behavior